In this paper a strategy to identify and select the most relevant variables to study problems in the exact sciences, when large databases of data have to be explored, is formulated. It consists of a first exploratory stage, performed mainly with the Classification and Regression Tree method, to determine the list of most relevant signals to be used in the analysis of the phenomenon of interest. A linear followed by a non-linear correlation technique (Principal Component Analysis and Auto-associative Neural Networks respectively) are then applied to reduce the number of signals to the ones containing non redundant information. The potential of the approach is illustrated by an application to the problem of identifying the confinement regime in the Joint European Torus. The minimum set of signals has been used to train a neural network and its performance is compared with various theoretical models. The success rate of the neural network is very high and it significantly outperforms the theoretical models in terms of classification accuracy.